Movement is an essential aspect of our lives. Not only do we move to interact with our physical environment, but we also express ourselves and communicate with others through our movements. In an increasingly computerized world where various technologies and devices surround us, our movements are essential parts of our interaction with and consumption of computational devices and artifacts. In this context, incorporating an understanding of our movements within the design of the technologies surrounding us can significantly improve our daily experiences. This need has given rise to the field of movement computing – developing computational models of movement that can perceive, manipulate, and generate movements. In this thesis, we contribute to the field of movement computing by building machine-learning-based solutions for automatic movement generation. In particular, we focus on using machine learning techniques and motion capture data to create controllable, generative movement models. We also contribute to the field by creating datasets, tools, and libraries that we have developed during our research. We start our research by reviewing the works on building automatic movement generation systems using machine learning techniques and motion capture data. Our review covers background topics such as high-level movement characterization, training data, features representation, machine learning models, and evaluation methods. Building on our literature review, we present WalkNet, an interactive agent walking movement controller based on neural networks. The expressivity of virtual, animated agents plays an essential role in their believability. Therefore, WalkNet integrates controlling the expressive qualities of movement with the goal-oriented behaviour of an animated virtual agent. It allows us to control the generation based on the valence and arousal levels of affect, the movement's walking direction, and the mover's movement signature in real-time. Following WalkNet, we look at controlling movement generation using more complex stimuli such as music represented by audio signals (i.e., non-symbolic music). Music-driven dance generation involves a highly non-linear mapping between temporally dense stimuli (i.e., the audio signal) and movements, which renders a more challenging modelling movement problem. To this end, we present GrooveNet, a real-time machine learning model for music-driven dance generation.
Copyright is held by the author(s).
This thesis may be printed or downloaded for non-commercial research and scholarly purposes.
Supervisor or Senior Supervisor
Thesis advisor: Pasquier, Philippe
Member of collection